{"id":"be17315c-7f4a-44aa-ac47-edb008039121","arxiv_id":"2411.19474","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A diffuse, wide-field LiDAR paired with RGB and an adaptive loss improves 3D reconstruction over sparse LiDAR in low-texture and low-light scenes.","lead":"This paper combines a blurry, wide-field time-of-flight sensor (diffuse LiDAR) with a regular camera to make 3D scans of objects and rooms. The authors show that, with few camera views and in dark or textureless scenes, this blurry depth sensor can beat the sparse dot-projecting LiDAR found in phones.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Comparative claim uses an underpowered 8x8 sparse-LiDAR baseline; realistic sparse-LiDAR densities may overturn the 'diffuse LiDAR outperforms' conclusion.","rationale":"The reader's weakest assumption was pose availability from RGB, which is a legitimate and explicitly acknowledged scope limitation, especially for the no-texture synthetic scenes where COLMAP cannot provide poses. However, the more load-bearing issue for the central claim is the representativeness of the sparse-LiDAR baseline. The paper's own stated mechanism—sparse LiDAR struggles with limited-view scene coverage—is reasonable, but an 8x8 point grid is an extreme instantiation of that limitation; real handheld sparse LiDARs can emit thousands of dots or provide dense depth maps. The rank analysis in Sec. 3 depends on the assumed number of sparse measurements, so the analytical support is also parameter-dependent. No error bars are reported, and real-world comparisons are qualitative with a simulated sparse baseline. These issues do not refute the method, but they make 'surprisingly outperform traditional sparse LiDAR' broader than the current evidence supports. The reader's CONDITIONAL verdict is therefore appropriate, and my specific concern does not move it; it should, however, be a required condition: report density-matched and realistic sparse-LiDAR comparisons before making the comparative claim.","tokens_in":13942,"tokens_out":8037,"duration_ms":73521,"concrete_test":"Re-run the Table 1 simulated experiments and the Fig. 3 rank analysis with sparse-LiDAR grids of 16x16, 32x32, and 64x64 points per view (and, if feasible, a commercial-grade sparse LiDAR point cloud at its native density), holding all other settings and pose inputs fixed. Plot depth/normal MAE and rank versus sparse-LiDAR density. If at any realistic density the sparse-LiDAR baseline matches or beats the diffuse-LiDAR method, the central comparative claim is not supported; if the diffuse method remains ahead across densities, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim ('diffuse LiDAR can outperform traditional sparse LiDAR') is established primarily in Table 1 and the rank analysis in Sec. 3/Fig. 3. In Sec. 5.1, the sparse-LiDAR baseline is explicitly '8x8 points located at the center of each diffuse LiDAR pixel zone' (Table 1). This equips the sparse baseline with one depth point per large diffuse-LiDAR pixel, not with the pixel/dot density of commercial handheld sparse LiDARs. The recoverability analysis in Sec. 3 likewise fixes the number of measurements, so it does not separate sensor modality from spatial sampling density. If a sparse LiDAR were given a denser grid (e.g., 32x32 or 64x64, as real devices can provide), its coverage and matrix rank would increase, and the observed depth/normal advantages (e.g., Tab. 1d D.MAE ~0.04 vs ~0.11) may shrink or reverse. The paper does not report a sensitivity analysis over sparse-LiDAR point density, and the real-world sparse baseline is simulated by subsampling a depth map, with no quantitative real comparison. Thus the headline claim is conditional on an unstated and unrepresentative choice of baseline density.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diffuse-LiDAR-plus-RGB framework for handheld 3D scanning, targeting low-texture, low-light, and low-albedo scenes. The authors render transient histograms from a wide-IFOV SPAD sensor in a Gaussian-surfel representation, and introduce a scene-adaptive loss that down-weights RGB where texture or SNR is low and up-weights the LiDAR transient loss. They claim that diffuse LiDAR outperforms conventional sparse point LiDAR when the number of views is limited, supported by a rank-based recoverability analysis, synthetic evaluations on four scenes with four texture variations, and qualitative real-world captures.","tokens_in":14136,"tokens_out":5540,"duration_ms":43028,"significance":"If the comparative claim survives stronger validation, the work is significant: it challenges the assumption that pointwise sparse LiDAR is the preferred depth modality for handheld fusion, and it provides a differentiable transient rendering pipeline and an adaptive loss that could be reused in other sensor-fusion settings. The recoverability analysis and the LiDAR-only ablation (Table 2) are useful contributions. However, the current evidence is weakened by the closed-loop synthetic evaluation, the underpowered 8x8 sparse-LiDAR baseline, and the absence of quantitative real-world comparison, so the practical significance is not yet fully demonstrated.","major_comments":[{"comment":"The simulated experiments in Sec. 5.1 render RGB and 8x8 diffuse LiDAR histograms using the image formation model described in Sec. 4 (Eqs. 10-12), and then evaluate the reconstruction against that same model. This makes the quantitative gains partly a measure of internal consistency with the assumed forward model, not of external validity. The real-world experiments in Sec. 5.2 provide only qualitative comparisons, and the sparse-LiDAR baseline is simulated by subsampling D435i depth maps, so they do not break the circularity. Please add either a real sparse-LiDAR sensor as a baseline or a quantitative validation of the forward model against the real hardware, and include a sensitivity analysis over forward-model parameters.","section":"Sec. 5.1, Eqs. (10)-(12)"},{"comment":"The sparse LiDAR baseline is explicitly '8x8 points located at the center of each diffuse LiDAR pixel zone' (Sec. 5.1). This gives the sparse baseline only 64 depth points per view, which is far below the density of commercial handheld LiDARs. The recoverability analysis in Sec. 3.2 also fixes the number of measurements, so it does not separate sensor modality from spatial sampling density. As a result, the headline claim that diffuse LiDAR outperforms sparse LiDAR is not established for representative sparse-LiDAR densities. Please report results for denser sparse grids (e.g., 16x16, 32x32, 64x64) in Table 1 and Fig. 3, and identify the density at which the comparison changes.","section":"Sec. 5.1, Table 1"},{"comment":"The paper states in Sec. 4 that COLMAP is used to obtain the camera intrinsics and per-frame extrinsics, but Table 1d presents results on completely textureless scenes, where COLMAP cannot estimate poses. The manuscript does not disclose the pose source for these experiments. Since the claimed robustness in no-texture scenes is a central contribution, please specify how poses were obtained for each dataset (e.g., ground-truth camera trajectories, known synthetic poses, or a LiDAR-based pose estimator) and discuss whether the method remains applicable when RGB-based pose estimation fails.","section":"Sec. 4, Table 1d"},{"comment":"The recoverability analysis in Sec. 3.2 relies on an approximate linear model y = Ax, with details deferred to the Supplementary, which is not included. The rank claims in Fig. 3 are load-bearing for the motivation, so the model should be specified in the main text or an available supplement, and its relationship to the nonlinear transient rendering model in Eqs. (10)-(12) should be justified. In addition, the statement that the analysis is performed in 2D 'without loss of generality' needs justification, since the rank behavior may differ in 3D.","section":"Sec. 3.2, Fig. 3"}],"minor_comments":[{"comment":"The caption says 'as RGB input SNR increases' but the text describes relying more on diffuse LiDAR as SNR decreases; the caption appears to have the direction reversed.","section":"Fig. 6 caption"},{"comment":"The sigmoid hyperparameters a, b, and k are not reported anywhere; please provide their values and note whether they are fixed across all experiments.","section":"Eq. (13), Sec. 4.2"},{"comment":"The table does not define the units of D.MAE and N.MAE; please specify (e.g., meters for depth, degrees for normals).","section":"Table 1"},{"comment":"The paper claims low-bandwidth advantages for diffuse LiDAR, but no concrete bandwidth comparison (e.g., number of histogram bins versus number of LiDAR points) is quantified; please provide a representative comparison.","section":"Sec. 3.1"},{"comment":"The RGB-LiDAR extrinsics calibration procedure is not described; please explain how the relative pose between the two sensors was obtained for the real captures.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising idea and a reasonable initial framework, but the central comparative claim currently rests on a circular synthetic evaluation and an underpowered sparse-LiDAR baseline. The missing pose-disclosure for the no-texture experiments is also a correctness concern. These issues are fixable within the scope of a revision, so I recommend major revision rather than rejection. I would also encourage the editors to ask for the supplementary material to be included in the review package, since the linear model details are deferred there."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper's real novelty is in putting diffuse (flash) SPAD ToF sensors together with RGB in a multi-view Gaussian-surfel reconstruction, with a per-patch adaptive loss that weighs RGB by local texture and SNR. I'm not aware of prior work doing exactly that, and the idea is plausible: the transient histograms from a wide-IFOV pixel are a natural fit for analysis-by-synthesis, and the adaptive weighting is a sensible way to let LiDAR take over where RGB has no signal. The recoverability analysis in Fig. 3 gives a nice intuition, and the qualitative results with the low-cost TMF8828 show the hardware is real.\n\nBut the quantitative case has soft spots. The synthetic evaluation renders scenes using the same forward model (Eqs. 10-12) that the reconstruction optimizes. That measures internal consistency, not external validity. The sparse-LiDAR baseline is 8x8 depth points per view. That's an underpowered baseline: it is matched to the diffuse sensor's 8x8 zones, but the paper's conclusion is stated broadly against 'traditional sparse LiDAR,' and real handheld LiDARs deliver more points. The rank analysis fixes the measurement count, so it doesn't separate sensor modality from sampling density. A simple sensitivity sweep over sparse point density would settle the issue, and it's missing.\n\nThe pose assumption is worth flagging too. The paper assumes COLMAP pose from RGB; in the no-texture synthetic experiments COLMAP can't work, so ground-truth poses must be in play without saying so. That makes the robustness claim conditional on pose availability, which is admitted in Scope but not in the abstract. Also, the real-world experiments are qualitative only, the real sparse baseline is simulated by subsampling the active stereo depth, and the adaptive-loss hyperparameters (a,b,k) aren't reported. No error bars anywhere. These are not fatal, but they need addressing.\n\nOverall, the paper is worth engaging with. The core idea is good, the writing is clear, and the authors are honest about limitations. A serious referee could reasonably ask for major revision: denser sparse-LiDAR baselines, a real LiDAR comparison, error bars, and ideally some quantitative real-world numbers. I'd send it to review, and I'd probably cite it if I worked on sensor fusion for 3D scanning.","headline":"Diffuse LiDAR plus RGB in a surfel framework is a fresh and plausible idea, but the headline comparison to sparse LiDAR rests on a weak 8x8 baseline and a circular synthetic evaluation.","tokens_in":14768,"tokens_out":3488,"would_cite":true,"duration_ms":30244,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Diffuse (blurred) LiDAR combined with RGB can reconstruct 3D scenes more accurately than conventional point-grid LiDAR when only a few views are available.","keywords":["diffuse LiDAR","transient histograms","Gaussian surfels","3D reconstruction","sensor fusion","scene-adaptive loss","handheld scanning","limited views"],"falsifier":"On a fixed textured-object/textureless-plane scene, sweep the number of training views from 2 to 30 and record depth MAE for RGB plus sparse LiDAR versus RGB plus diffuse LiDAR; if sparse LiDAR reaches parity at or below the 10-view regime, the claimed limited-view advantage does not hold.","tokens_in":13630,"feed_emoji":"🔦","tokens_out":7704,"duration_ms":61881,"temperature":0.7,"pith_summary":"This paper argues that a 'blurred' LiDAR—one that emits a diffuse flash and records each pixel as a time-of-flight histogram spanning a wide field of view—can beat the sparse point-grid LiDARs now found in phones, provided it is fused with RGB. The reason is coverage: a wide-field pixel captures many scene points at once, so a coarse 8 by 8 sensor can supervise geometry across an entire scene, while a sparse LiDAR leaves gaps that never get filled when only a handful of views are available. The authors build a Gaussian-surfel renderer that synthesizes both RGB images and transient histograms, and a scene-adaptive loss that leans on RGB where it has texture and signal, and on LiDAR where it does not. If the claim holds, handheld 3D scanning becomes more reliable exactly in the conditions that break camera-only reconstruction: low texture, low light, and low albedo.","feed_headline":"Blurred LiDAR beats sparse LiDAR for handheld 3D scans","feed_subtitle":"Paired with RGB and an adaptive loss, a low-res diffuse flash fills depth gaps that point-grid LiDAR leaves open.","key_machinery":"The engine of the method is a Gaussian surfel scene representation—surfaces built from flattened, oriented 2D Gaussians—rendered through differentiable alpha-compositing into both RGB images and transient histograms. Transient rendering works by sampling rays within each wide-IFOV LiDAR pixel, projecting each surfel's depth into a time bin, and soft-binning the contribution so gradients flow. Around this renderer sits a scene-adaptive loss: each RGB patch is scored by texture variance and signal-to-noise ratio, and the final loss weights RGB L1 plus SSIM by that score while weighting the transient KL divergence by its complement. This per-patch reweighting is what lets the optimizer dynamically choose the sensor that is informative at each location, which is the mechanism that converts diffuse LiDAR coverage into sharper depth, normals, and meshes.","core_discovery":"The paper's central discovery is that the spatial blur of diffuse LiDAR is not a defect to be removed but a coverage advantage to be exploited. A diffuse LiDAR pixel integrates time-of-flight returns from every surface in its instantaneous field of view, so a low-resolution 8 by 8 sensor still sees the whole scene; the mixed returns are disentangled by rendering predicted transients from a Gaussian surfel scene and comparing them to measured histograms. RGB supplies the dense spatial and color structure that resolves the per-pixel depth ambiguity. A scene-adaptive loss computes per-patch texture variance and signal-to-noise ratio to decide, patch by patch, whether RGB or LiDAR should dominate: textured high-SNR patches rely on RGB, while textureless, dark, or low-albedo patches fall back on LiDAR. In simulated scenes at 10 training views, this combination cuts depth error roughly in half or better relative to RGB plus sparse LiDAR and, in no-texture scenes, reaches depth MAE around 0.04 versus roughly 0.11 for sparse LiDAR; real captures at 90 views show improved object-plane separation. The paper also shows analytically that increasing LiDAR pixel field of view raises the rank of the linear measurement operator when views are scarce, which is why blur helps precisely in the few-view regime.","pith_inferences":["The paper does not test this, but the coverage advantage should shrink as view count grows; a view-count sweep would reveal the crossover at which sparse LiDAR's precise points catch up.","The scene-adaptive weights could double as a per-patch confidence map, guiding downstream tasks that need to know whether the geometry is backed by RGB or LiDAR evidence.","Because pose estimation currently comes from RGB alone, extending the framework to use diffuse LiDAR transients for pose initialization or refinement would attack the main remaining failure mode in fully textureless scenes."],"forward_implications":["In limited-view scans with 10 training views, diffuse LiDAR improves depth and normal estimation over sparse LiDAR across full, mixed, and no-texture scenes, with no-texture depth MAE roughly three times lower.","The scene-adaptive loss automatically down-weights RGB as its SNR degrades, so depth estimation stays stable from high light down to noise levels that fully obscure RGB cues.","Diffuse LiDAR alone, without RGB, beats both sparse point LiDAR and sparse histogram LiDAR under the same low-resolution capture setup.","On real handheld captures with 90 views, the method improves object-plane separation and geometry in low-texture, low-light, low-albedo settings.","The recoverability analysis shows that wider IFOV raises the linear measurement matrix rank when views are scarce, so the benefit is specifically a limited-view phenomenon."],"supporting_citations":[{"why":"Supplies the Gaussian surfel representation and the depth-normal regularization that carries the reconstruction.","marker":"[5]"},{"why":"Provides the alpha-compositing rasterization and spherical-harmonic color model used for RGB rendering.","marker":"[15]"},{"why":"Supplies the transient-radiance-field rendering and soft histogram binning that the diffuse LiDAR rendering builds on, and serves as a sparse-histogram baseline.","marker":"[23]"},{"why":"Provides the structure-from-motion pipeline assumed to give camera intrinsics and per-frame poses for all experiments.","marker":"[33]"},{"why":"Defines the standard sparse-LiDAR depth supervision baseline that the paper's fused method is compared against.","marker":"[6]"},{"why":"Supplies the monocular depth prior baseline that is expected to fail on low-texture scenes.","marker":"[44]"},{"why":"Documents the low-cost 8 by 8 SPAD sensor used for real-world diffuse LiDAR captures.","marker":"[1]"},{"why":"Supplies the view-number analysis that motivates why wider field of view improves recoverability with few views.","marker":"[36]"}],"fun_headline_variants":["Blurred LiDAR beats sparse grid for few-view 3D scans","Diffuse flash LiDAR outperforms point-grid LiDAR in dark scenes","Why a blurry LiDAR is better for handheld 3D scanning","RGB and diffuse LiDAR combine for robust 3D scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline assumes that accurate per-frame camera poses can be obtained from RGB alone (via COLMAP), which may fail in the textureless and low-light scenes the method targets.","fun_headline_variants_meta":{"raw":{"variants":["Blurred LiDAR beats sparse grid for few-view 3D scans","Diffuse flash LiDAR outperforms point-grid LiDAR in dark scenes","Why a blurry LiDAR is better for handheld 3D scanning","RGB and diffuse LiDAR combine for robust 3D scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000563,"raw_usage":{"total_tokens":2707,"prompt_tokens":1018,"completion_tokens":1689,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":1611}},"tokens_in":634,"tokens_out":1689,"duration_ms":13402,"temperature":1.0,"reasoning_tokens":1611,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:09:42.873719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fixed textured-object/textureless-plane scene, sweep the number of training views from 2 to 30 and record depth MAE for RGB plus sparse LiDAR versus RGB plus diffuse LiDAR; if sparse LiDAR reaches parity at or below the 10-view regime, the claimed limited-view advantage does not hold.","supporting_citations":[{"cited_title":"High-quality surface recon- struction using gaussian surfels","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian surfel representation and the depth-normal regularization that carries the reconstruction."},{"cited_title":"Transient neural radiance fields for lidar view synthesis and 3d reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the transient-radiance-field rendering and soft histogram binning that the diffuse LiDAR rendering builds on, and serves as a sparse-histogram baseline."},{"cited_title":"Structure- from-motion revisited","cited_arxiv_id":null,"evidence_quote":"Provides the structure-from-motion pipeline assumed to give camera intrinsics and per-frame poses for all experiments."},{"cited_title":"Depth anything: Unleashing the power of large-scale unlabeled data","cited_arxiv_id":null,"evidence_quote":"Supplies the monocular depth prior baseline that is expected to fail on low-texture scenes."},{"cited_title":"TMF882X Datasheet","cited_arxiv_id":null,"evidence_quote":"Documents the low-cost 8 by 8 SPAD sensor used for real-world diffuse LiDAR captures."},{"cited_title":"Role of tran- sients in two-bounce non-line-of-sight imaging","cited_arxiv_id":null,"evidence_quote":"Supplies the view-number analysis that motivates why wider field of view improves recoverability with few views."}],"review_version":1}